The Model Pricing card controls the estimated cost shown on the AI Transparency page. SOBS stores pricing per model as USD per 1 million input and output tokens, then applies those values to each visible AI span to calculate the Est. Cost summary and row-level estimates.

• Default rows are built into SOBS for common OpenAI, Anthropic, Gemini, Llama, and Mistral models.
• Inferred rows appear automatically when SOBS sees a model name in telemetry that is not already configured. The row is seeded from the closest known family, such as gpt-4o*, claude sonnet/haiku/opus, gemini flash/pro, or llama.
• Custom rows are for models SOBS cannot predict or for internal aliases. Click Add Custom Model, type the model name in the new row, then enter input and output pricing.
• Changing any price updates future cost estimates immediately after saving. Existing telemetry is not rewritten; only the display estimate changes.

1. SOBS scans AI trace data for distinct values in gen_ai.request.model.
2. If a model already has a built-in or saved pricing row, that value is used.
3. If the model is new, SOBS seeds a row from the closest known family so cost cards remain useful instead of blank.
4. You should review inferred rows for provider-specific pricing differences, preview models, or internal deployment aliases.

The inferred badge is only a starting signal. Once you know the real rate, edit the row and save it. SOBS persists the saved value under that exact model name.

Clicking Add Custom Model inserts a new editable row at the bottom of the pricing table. The first field in that row is the model name. Enter the exact model string you expect to see in telemetry, for example my-company-reasoner-v2, then provide the input and output token prices.

• Use the exact emitted model name when possible so the estimate matches without relying on fuzzy family matching.
• Custom rows can be removed before saving with the delete button on the right.
• Default and inferred rows keep the model name locked because they are keyed off known or observed telemetry values.

• Helper widget responds on at least one telemetry page.
• Guard model blocks intentionally unsafe test prompt.
• Token validation status shows healthy/valid state.
• Repository owner/repo string resolves for agent issue creation.
• If DLP is enabled, endpoint returns a JSON flagged boolean.
• AI Transparency shows cost values that look reasonable for your primary models after saving pricing changes.

• If helper appears disabled, check endpoint URL and API key first.
• If all AI actions fail, verify guard endpoint/model availability.
• If GitHub actions fail, validate token scopes and expiration date.
• If assignments stall, review per-hour and active-assignment limits.
• If a cost looks wrong, compare the telemetry model name to the pricing row key first. A custom row using the exact model string is the safest override.
• If a new model appears with an inferred badge, treat that as a prompt to verify the actual provider pricing and update it if needed.